I like to end my work weeks by reading an academic paper on AI, scintillating I know. 🙂 However, I find that when a subject is changing quickly or if there is a lot of hype, it is often best to go to the source material to know exactly what is going on. With computer science we have the added advantage of being able to get the corresponding code from GitHub to try it out. In business you have to spend extra time scrutinizing the SEC filings to normalize the financials for comparison. I have been impressed with the results from deep learning and have been excited about what should be possible in the future. Since I feel this way, it is time for me to seek out an alternative perspective. So I recently read “Deep Learning: A Critical Appraisal” by Gary Marcus who is a professor in the Department of Psychology and Neural Science at NYU. It is an interesting paper and was controversial when it was published. I will give a brief review of the paper and a few of my thoughts.
To get everyone on the same page, deep learning is a statistical model for classifying data and is at the core of many of the AI advances we enjoy today. Labeled data is used to train a model and when a new piece of information is presented it classifies it based on data that it was previously exposed to.
Even with the great many applications that this technique has there are 10 challenges Gary Marcus calls out:
- Deep learning is data hungry and requires too much data to make basic classifications that humans can often do with a single example.
- Despite its name, deep learning is only deep architecturally and not from a knowledge perspective with limited transferring of insights.
- There is no natural way to deal with hierarchical structure. Changing the order of words leaving a sentence open to interpretation can be difficult when the generalized set differs widely from the training set.
- Deep learning struggles with open-ended inference. It should be noted that this is a problem that the industry is trying to address through new AI contests like SQuAD v2.
- People are concerned with the ethics of AI and the lack of transparency does not help. As AI becomes more central to a lot of decision making it will be critical that the models can be audited for the right level of accountability.
- Prior knowledge has not been well integrated into deep learning models. Of the objections this is one that I personally enjoyed. Models are interpretations based on previously seen data but there are some bodies of knowledge like physics that have formulas that can be used explicitly on new examples without requiring interpretation.
- Causation and correlation are not inherently differentiated with deep learning.
- The world is largely assumed to be stable and it is not. That is why today AI is great at playing games but is still difficult to use in open-ended problem spaces.
- Deep learning is great at approximation but cannot be fully trusted. This reminds me of using language translation where the results are really good for establishing understanding but at this time I would still want a human if I was dealing with a contract.
- It is difficult to engineer deep learning. Although true, there is a lot to disagree with. The frameworks (e.g. PyTorch, TensorFlow, Keras, ONNX, etc.) are making it easier to have common building blocks that takes some of the difficulty out of development including supporting popular neural network structures including new ones like generative adversarial networks (GANs).
The paper highlights ten challenges but also includes 4 ways to make it better that I will expand upon here.
Paper’s Recommendations for AI
Unsupervised learning is a theoretically better approach for learning which is something that I agree with. I wrote a news reader that I use to provide a conservative and liberal perspectives to news stories. Problems with labeling data in sufficient quantity to make this work is tough. Instead I used document clustering so that I would not have to rely on labeled data. This strategy worked really well in providing me a more balanced news reader. With that said unsupervised learning still has further to go if it is going to surpass supervised learning.
Symbol-manipulation, and hybrid models
Hybrid models can be used in addressing the prior knowledge problem that was raised as part of the challenges. By encoding math, physics and other codified knowledge into a hybrid model there could be definitive knowledge using best practices for getting answers and then use the deep learning process for classifying information when there is not a symbol-manipulation method to do it. Note that in the critiques of this paper this was an area aggressively pushed back on since the results historically haven’t been as good as deep learning. Although true, there was a time when it was said that deep learning was a failure and now it is the dominant AI approach so it might not be right to completely dismiss symbol-manipulation.
More insights from cognitive development
Ethics is something that is top of mind for people dealing with AI and trying to make sure that it has a positive impact on society. Part of the way to deal with these issues is to not just rely on mathematics but to also include insights from human psychology. In principle I really like this concept but I don’t know enough about psychology to determine how much of it can be integrated into models. There is also the notion of common sense being integrated into the models. However, looking at the many issues being publicly discussed it is clear that there would be a lot of problems in determining what exactly is “common” in our collective sense. Math and science would be a lot easier to research and determine than to add human psychology into the mix but this would be an important development.
AI contests have done a great job of rallying industry and academia to take on new challenges. However, if the goal is arguably to reach artificial general intelligence, then there needs to be bolder challenges that expand beyond confined closed-end problems. There is some signs that this is becoming a thing accepted in the AI community. For example, the Stanford Question and Answer Dataset (SQuAD) originally started by finding answers in a passage but in the second version this has been updated to not only find answers in a passage but also to determine if something is not in the passage. This is a model that should be replicated in a lot of the tests that are out there.
Overall this was a good read with a lot of thought provoking questions. There is a lot of great AI advances with more progress to come. Advances in computing from processors (CPUs to GPUs to FPGAs) will only do so much but advances in quantum computing should provide a fundamental step function in making the existing deep learning process more effective. However, even with that there will most likely need to be a new approach to getting to the next level like what happened in the past from the transition to neural networks from expert systems. Geoff Hinton mentioned this as well with “science progresses one funeral at a time. The future depends on some graduate student who is deeply suspicious of everything I have said.”
Talk to you later,