This is written in response to a Quora question, which asks about the street value of pre-trained models. Feel free to vote there for my answer! .

This is an interesting question. There’s clearly no definitive answer to it. My personal impression (partly based on my own experience) that with a reasonable amount of training data, pre-training and/or data augmentation is not especially useful (if at all). In particular:

1. In a recent paper by Facebook, this is demonstrated for an image-detection/segmentation task: Re-thinking ImageNet pre-training. He et al. 2018.
   
2. A couple of recent chilling results:
   1. Researchers from Google and Carnegie Mellon university showed that a 300x (!) increase in the number of training examples only modestly improves performance. I think it is an especially interesting result, because the data is only weakly supervised (i.e., it is the most realistic big-data scenario).
   2. Unsupervised training does not work yet for truly low-resource languages: Two New Evaluation Data-Sets for Low-Resource Machine Translation: Nepali–English and Sinhala–English. Guzman et al 2018.
3. Here is one example from the speech-recognition domain: Exploring architectures, data and units for streaming end-to-end speech recognition with RNN-transducer. Rao et al 2018, where pre-training works, but gains are rather modest: "We find CTC pre-training to be helpful improving WER 13.9%→13.2% for voice-search and 8.4%→8.0% for voice-dictation".

Thus, if you are interested in obtaining SOTA results on the dataset of interest, you may need to be very clever and efficient in obtaining tons of training data. That said, pre-training certainly allows one to achieve better results in many cases, especially when the amount of training data is small. This can be really useful for bootstrapping. See, e.g., the following radiology paper: Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. Shin Hoo-Chang et al. 2016.

That said, AI is a fast-developing field, and we see particularly impressive advances in transfer learning for NLP. This series largely started with the following great papers (particularly from the Allen AI ELMO paper):

1. Semi-supervised Sequence Learning. Andrew M. Dai, Q. Le. 2015.
2. ELMO paper: Deep contextualized word representations. Peters et al 2018.

Recently, we have seen quite a few improvements on this with papers from Open AI (GPT), Google AI (BERT), and Microsoft (I think it’s called Big Bird, but I am a bit uncertain). These improvements are huge and very encouraging. Let us not forget that the road to these successes have been paved by two seminal papers, which largely started the neural NLP:

1. Natural language processing (almost) from scratch. R Collobert, J Weston, L Bottou, M Karlen. 2011.
2. Distributed representations of words and phrases and their compositionality. 2013. T Mikolov, I Sutskever, K Chen, GS Corrado, J Dean.

Both papers proposed its own variant of neural word embeddings, learned in an unsupervised fashion. This was clearly a demonstration of the street value of a pre-trained model in NLP. Furthermore, the first paper, which was a bit ahead of the time, went much further and presented possibly the first suit of core neural NLP tools (for POS tagging, named entity recognition, and parsing). It is worth mentioning that there are also earlier and less-known papers on neural NLP including (but not limited to) a seminal neural language modeling paper by Y. Benigo.

In conclusion, I would also note that a lot of pre-training has been done in the supervised fashion. Perhaps, this is a limiting factor as the amount of supervised data is relatively small. We may be seeing this changing with more effective unsupervised pre-training methods. This has become quite obvious in the NLP domain. However, there is a positive trend in the image community too. For example, in this recent tutorial (scroll to the unsupervised tutorial pre-training), there is a couple of links to recent unsupervised training approaches that rival ImageNet pre-training.

Some further reading: A good overview of the transfer learning is given by S. Ruder.

This is written in response to a Quora question, which asks if most NLP engineers will be out of jobs once computers are capable of near-perfect text and speech processing. Feel free to vote there for my answer on Quora! There is also an interesting recent blog post by a Carnegie Mellon professor Jeffrey P. Bigham, where he expresses similar concerns and argues that in the near future artificial intelligence should be treated as a human-computer interaction problem rather than purely an algorithmic problem. I fully agree with this point of view!

So, will NLP be engineers be out of jobs? Yes, absolutely! Future is highly uncertain. As believed by Marvin Minsky, an imperfect human race will create a new robotic race that will not suffer from human limitations and which will inherit the Earth and surrounding planets. As humans have outcompeted and replaced many other species, these robots will outcompete and replace humans. We should not, however, fear the future but fulfill our evolutionary destination and welcome our robot overlords.

As impressive as they are, existing AI systems only seem to be intelligent. We do not know how many dozens, hundreds, and, possibly, thousands of years it will take to create truly intelligent machines. In fact, we do not truly know what it means to be intelligent and what is required to be intelligent. A recent high-profile paper: “Building Machines That Learn and Think Like People”, 2016 by Lake et al., tries to find some answers, but its conclusions are far from being definitive.

Ray Kurzweil famously predicted singularity to happen in 2045 based on the exponential growth of computational capacity. However, the size of the transistor is already about 100x the size of an atom. My guess would be that the current technology has a potential of a 10x increase in capacity. It also seems that there is no production-ready immediate replacement on the horizon. In particular, it is not clear when (and if) 3-d chips will be available.

At the same time, the best GPUs have about 20 billion transistors, while the human brain has 100 billion neurons each of which has 10K connections (synapses) on average. How many transistors are necessary to create an artificial neuron? One of the most advanced custom neural chips TrueNorth implements one million spiking neurons and 256 million synapses on a chip with 5.5 billion transistors with a typical power draw of 70 milliwatts.

Thus, it takes about 20 transistors per synapse. Even if we assume that an artificial neuron is as powerful as the real one (which is likely very far from truth), the current technology is six freaking orders of magnitude behind a human brain! Size notwithstanding, power consumption is also an enormous challenge. According to the above cited report, if TrueNorth is scaled up to the size of the human brain it would require 10,000 times more energy!

Furthermore, it is highly unrealistic to assume that an artificial neuron is nearly as complex as a real one. For example, the following book argues that even a single-cell organism (albeit a rather large one) can exhibit extremely complex behaviors, which include sensing and hunting: Wetware: A Computer in Every Living Cell: Dennis Bray. C elegans has about 500 neural cells, but it has basic sensory system and muscle control! It can reproduce and mate.

As my co-author and friend Daniel Lemire noted, our planes do not fly like birds and submarines do not swim like fishes. We do not have to mimic human brain to solve artificial intelligence tasks. We may not even need a brain-like structure to create a truly thinking machine. However, I would argue that we—using the phrase of the Turing award winner Richard Hamming—simply do not have an attack, i.e., a reasonable way to approach this difficult problem.

Another good observation from Daniel Lemire is that we should expect the unexpected because experts can be easily wrong. For example, there were some predictions about impossibility of flight in early 20th century. Although we should expect breakthroughs anytime, I do not think that impossibility of flight for heavier than air machines is a good example. The first gliders appeared well before the first propelled planes. In fact, some people had very clear ideas about how planes should and could fly. This is not true for the general artificial intelligence and we have not built the first gliders yet.

Traveling to the stars is clearly a difficult problem. Few people would argue with that. However, for some reason everybody thinks that artificial intelligence is something that is just a few (dozens) years away. Well, it could be so. But it could also be harder than interstellar travel.

Even if we can create a human-size neural network, we do not know how to program it efficiently. A state-of-the-art approach to training a model consists in collecting a huge amount of data and making a neural network that finds a mapping from inputs to outputs. This approach truly revolutionizes speech and vision and improves to some degree text processing. However, it might be just a gigantic “fuzzy” memory.

This approach is also incredibly brittle and data greedy. We do not know if we can scale it from hundreds to millions of layers. There are a number of recent papers showing that it is very easy to “poison” training data. For example, in the IMDB sentiment dataset the error rate can be driven from 12% to 23% by adding only 3% poisoned data. State-of-the-art CNNs fail (accuracy drops from 90+% to 10-%) for color modified CIFAR-10 images that are easily classified by humans.

Another big success of neural networks is speech recognition. Perhaps, it is the biggest success so far. For clean speech we can get near human recognition rates. However, on noisy data and especially when multiple speakers are present (an infamous cocktail party setting) the results are quite subhuman. The cocktail party setting is especially bad. It is a big success if you can reduce the word error rate from 90% down to 50% or to 30% (i.e., a computer misses every second or third word).

One clear issue with the current approaches is that clean training data can be quite expensive to obtain. For the existing not-so-clean data (collected in a semi-supervised fashion), there can be only little benefit by scaling (an already huge) training set by further two (!) orders of magnitude.

For example, a recent work by researchers from Google and Carnegie Mellon university has showed that a 300x (!) increase in the number of training examples only modestly improves performance. There is a lot of hope that reinforcement learning will solve these issues, but it does not seem to work yet.

All in all, judging by a good number of publications and blog posts that I have been reading in the last six years, we can now do well in a number of constrained domains. However, the success depends mostly on the existence of human-created training data and tons of engineering effort. In that, I suspect that the success of end-to-end systems (i.e., no engineering effort to modularize the problem and synthesize a system from multiple sometimes handcrafted models) is still limited.

Extending existing techniques to new domains requires many years of work from skilled engineers and scientists. I do not see how this can change in the near future. I actually expect that we will need many more scientists and engineers to continue making good progress. Brace yourself, it looks there is megatons of work ahead.

Last week I shared my time between work and hacking at a forth machine translation marathon in the Americas. This event organized jointly by CMU and Amazon (sponsored by Amazon) was a lot of fun. My small sub-team of two people got familiar with OpenNMT, trained English-German and English-Ukrainian models, as well as implemented an idea of our team lead Adithya Renduchintala. Hey, we have even gotten a tiny 0.5 gain in BLEU for the English-Ukrainian pair!

We certainly learned a lot of lessons, most of which generalize well beyond the neural machine translation task. One is related to implementation of custom neural modules and PyTorch. Unlike TensorFlow and many other packages, PyTorch belongs to a new crop of neural frameworks, where a neural network (computation) graph is dynamic. What does it mean? It means that you do not have define a computation graph in advance. You can simply write a tensor-manipulating code and PyTorch will do all the back-propagation and parameter updating automatically. Another well-known package with a similar functionality is DyNet.

There are ups and downs to dynamic computation graphs. For one thing, it is much simpler to debug them. For another, there is a lot of magic going behind the scenes, which you need to understand. First of all, one needs to remember that the computation graph is defined by a sequence of manipulations on Tensors and Variables (Variable is a Tensor wrapper that got deprecated in the recent PyTorch). Your sequences should be valid and properly linked so that all the Tensors of interest have a chance to be updated during back-propagation.

Tensor-level manipulations can easily get hairy. To simplify things, PyTorch introduces an abstraction layer called Module. A Module is a basic building block that has some parameters and a function forward to turn inputs to outputs. If all is done properly, given only the forward function PyTorch can compute the gradients via back-propagation and update the model parameters. The nice thing about PyTorch is that you can easily write a new module by combining several existing ones. There is no need to write arcane description of layers! As we can see from this PyTorch example, we can define a forward network computation in a very straightforward way. Even if you have not seen a line of PyTorch code before, you can easily figure out that this module applies two 2d-convolutions each of which is followed by a RELU non-linearity:

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)
 
    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

But here is a catch that is barely mentioned in PyTorch documentation. Although writing the forward function is sufficient to compute the gradients, it is apparently not sufficient to determine which tensors represent module's parameters. In the above example, the module includes two convolutional neural networks, each of which has parameters to be updated. How does PyTorch know this? Well, turns out that PyTorch overloads the function __setattr__! Thus, it surreptitiously "registers" each submodule when a user makes an assignment like this one:

self.conv1 = nn.Conv2d(1, 20, 5)

Unfortunately, such automatic registration does not work all the time. Imagine, for example, you want to aggregate several sub-modules whose number is not known in advance. It is very natural to save them all in a list:

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.sub_modules = []
        self.sub_modules.append(nn.Conv2d(1, 20, 5))  # append a module to the list
        self.sub_modules.append(nn.Conv2d(20, 20, 5)) # append a module to the list

Yet, this is where PyTorch magic stops: If you place the modules in plain Python list, PyTorch will not be able to update their parameters. As a fix, you need to register them explicitly. One way to do this is to explicitly call the function add_module. A less tedious ways is to use a combination of nn.ModuleList and nn.Sequential. Please, read a discussion here for more details.

The final version of my thesis "Efficient and Accurate Non-Metric k-NN Search with Applications to Text Matching" is now available online. An important by-product of my research is an efficient NMSLIB library, which I develop jointly with other folks. In a podcast with Radim Řehůřek (Gensim author) I discuss this project, its goals, and its history in detail.

Although efficiency is an important part of the thesis, it is primarily not about efficiency. Most importantly, I try to deliver the following messages:

1. We have very flexible retrieval tools, in particular, graph-based retrieval algorithms, which can work well for a wide variety of similarity functions. In other words, we do not have to limit ourselves neither to inner-product similarities (e.g., the Euclidean distance) nor to even metric spaces.
2. When "queries" are long, these algorithms can challenge traditional term-based inverted files. So, in the future, I expect retrieval systems to rely less on classic term-based inverted files and more on generic k-NN search algorithms (including graph-based retrieval algorithms). I think it is not a question of "IF", but rather a question of "WHEN".

Graph-based retrieval is an old new idea, which has been around for more than twenty years. This idea is beautifully simple: Build a graph where sufficiently close points are connected by edges. Such graphs come in various flavors and can be collectively called neighborhood graphs or proximity graphs. Given a neighborhood graph, nearest neighbor and other queries can be answered (mostly only approximately) by traversing the graph in a direction towards the query (and starting from, e.g., a random node). I cover the history of this idea in my thesis in more detail, but the earliest reference for this approach that I know is the seminal paper by Sunil Arya and David Mount (BTW, David Mount is co-author of the well-known ANN library).

Despite this early discovery, the practicality of graph-based retrieval in high-dimensional spaces was limited because we did not know how to construct neighborhood graphs efficiently. As it often happens in science, a number of fancy methods were proposed (while overlooking a simpler working one). Luckily, it was discovered that the graph can be constructed by iteratively building the graph and using a graph-based retrieval algorithm to find nearest neighbors for a new data point. A summit (or at least a local maximum) of this endevour is a Hierarchical Navigable Small World graph (HNSW) method, which combines efficient pruning graph-pruning heuristics, a multi-layer and multi-resolution graph topology with a bunch of efficiency tricks.

It was also known (but not well-known) that graph-based retrieval algorithms can work for generic (mostly metric) distances. So, I personally was interested in pushing these (and other) methods even further and applying them to non-metric and non-symmetric similarities. One ultimate objective was to replace or complement a standard term-based inverted file in the text retrieval scenario. Well, the idea to apply k-NN search to text retrieval is not novel (see, again, my thesis for some references). However, I do no think that anybody has shown convincingly that this is a viable approach.

On the way towards achieving this objective, there are a lot of difficulties. First of all, it is not clear which representations of text and queries one can use (I have somewhat explored this direction, but the problem is clearly quite hard). Ideally, we would represent everything as dense vectors, but I do not think that the cosine similarity between dense vectors is particularly effective in the domain of adhoc text retrieval (it works better for classification, though). I am also convinced that in many cases whenever dense representations work well, a combination of dense and sparse bag-of-words representations works even better. Should we embrace these hybrid representations in the future, we cannot use traditional term-based inverted files directly (i.e., without doing a simpler search with subsequent re-ranking). Instead, we are likely to rely on more generic algorithms for k-nearest neighbor (k-NN) search.

Second, instead of trying to search using a complex similarity, we can use such a similarity only for re-ranking. Of course, there should be obviously limits to the re-ranking approach. However, a re-ranking bag-of-words pipeline (possibly with some query rewriting) is a baseline that is hard to beat.

Third, k-NN search is a notoriously hard problem, which in many cases cannot be solved exactly without sequentially comparing the query with every data point (the so called brute-force search). This is due to a well-known phenomenon called the curse of dimensionality. Often we have to resort to using approximate search algorithms, but these algorithms are not necessarily accurate. How much inaccuracy is ok? From my experimental results I conclude that the leeway is quite small: We can trade a bit of accuracy for extra efficiency, but not too much.

Because approximate k-NN search leads to loss in accuracy, in my opinion, it does not make sense to use it with simple similarities like BM25. Instead, we should be trying to construct a similarity that beats BM25 by a good margin and do retrieval using this fancier similarity. My conjecture is that by doing so we can be more accurate and more efficient at the same time! This is one of the central ideas of my thesis. On one collection I got promising results supporting this conjecture (which is BTW an improvement of our CIKM'16 results). However, more needs to be done, in particular, by comparing against potentially stronger baselines.

In conclusion, I note that this work would have been impossible without encouragement, inspiration, help, and advice of many people. Foremost, I would like to thank my advisor Eric Nyberg for his guidance, encouragement, patience, and assistance. I greatly appreciate his participation in writing a grant proposal to fund my research topic. I also thank my thesis committee: Jamie Callan, James Allan, Alex Hauptmann, and Shane Culpepper for their feedback.

I express deep and sincere gratitude to my family. I am especially thankful to my wife Anna, who made this adventure possible, and to my mother Valentina who encouraged and supported both me and Anna.

I thank my co-authors Bileg Naidan, David Novak, and Yury Malkov each of whom greatly helped me. Bileg sparked my interest in non-metric search methods and laid the foundation of our NMSLIB library. Yury made key improvements to the graph-based search algorithms. David greatly improved performance of pivot-based search algorithms, which allowed us to obtain first strong results for text retrieval.

I thank Chris Dyer for the discussion of IBM Model 1; Nikita Avrelin and Alexander Ponomarenko for implementing the first version of SW-graph in NMSLIB; Yubin Kim and Hamed Zamani for the discussion of pseudo-relevance feedback techniques (Hamed also greatly helped with Galago); Chenyan Xiong for the helpful discussion on embeddings and entities; Daniel Lemire for providing the implementation of the SIMD intersection algorithm; Lawrence Cayton for providing the data sets, the bbtree code, and answering our questions; Christian Beecks for answering questions regarding the Signature Quadratic Form Distance; Giuseppe Amato and Eric S. Tellez for help with data sets; Lu Jiang for the helpful discussion of image retrieval algorithms; Vladimir Pestov for the discussion on the curse of dimensionality; Mike Denkowski for the references on BLUE-style metrics; Karina Figueroa Mora for proposing experiments with the metric VP-tree applied directly to non-metric data. I also thank Stacey Young, Jennifer Lucas, and Kelly Widmaier for their help.

I also greatly appreciate the support from the National Science Foundation, which has been funding this project for two years.