In an earlier blog post, I expressed my scepticism regarding the scientific value of non-transparent machine learning approaches, which only provide a result but no transparent explanation of how they arrive at their conclusion. I am aware that I run the risk of giving the impression of abusing this blog for my own agenda, against artificial intelligence and machine learning approaches in the historical sciences, by bringing the problem up again. However, a recent post in Nature News (Castelvecchi 2016) further substantiates my original scepticism, providing some interesting new perspectives on the scientific and the practical consequences, so I could not resist mentioning it in my post for this month.
Deep learning approaches in research on artificial intelligence and machine learning go back to the 1950s, and have now become so successful that they are starting to play an increasingly important role in our daily lives, be it that they are used to recommend to us yet another book that somebody has bought along with the book we just want to buy, or that they allow us to take a little nap while driving fancy electronic cars and saving carbon footprints for our next round-the-world trip. The same holds, of course, also for science, and in particular for biology, where neural networks have been used for tasks like homolog detection (Bengio et al. 1990) or protein classification (Leslie et al. 2004). This is true even more for linguistics, where a complete subfield, usually called natural language processing, has emerged (see Hladka and Holub 2015 for an overview), in which algorithms are trained for various tasks related to language, ranging from word segmentation in Chinese texts (Cai and Zhao 2016) to the general task of morpheme detection, which seeks to find the smallest meaningful units in human languages (King 2016).
In the post by Castelvecchi, I found two aspects that triggered my interest. Firstly, the author emphasizes that answers that can be easily and often accurately produced by machine learning approaches do not automatically provide real insights, quoting Vincenco Innocente, a physicist at CERN, saying:
As a scientist ... I am not satisfied with just distinguishing cats from dogs. A scientist wants to be able to say: "the difference is such and such." (Vincenco Innocente, quoted by Castelvecchi 2016: 22)This expresses precisely (and much more transparently) what I tried to emphasize in the former blog post, namely, that science is primarily concerned with the questions why? and how?, and only peripherally with the question what?
The other interesting aspect is that these apparently powerful approaches can, in fact, be easily betrayed. Given that they are trained on certain data, and that it is usually not known to the trainers what aspects of the training data effectively trigger a given classification, one can in turn use algorithms to train data that will betray an application, forcing it to give false responses. Castelvecchi mentions an experiment by Mahendran and Vedaldi (2015) which illustrates how "a network might see wiggly lines and classify them as a starfish, or mistake black-and-yellow stripes for a school bus" (Castelvecchi 2016: 23).
Putting aside the obvious consequences that arise from abusing the neural networks that are used in our daily lives, this problem is surely not unknown to us as human beings. We can likewise be easily betrayed by our expectations, be it in daily life or in science. This, finally, brings us back to networks and trees, as we all know how difficult it is at times to see the forest behind the tree that our software gives us, or the tree inside the forest of incompletely sorted lineages.
- Bengio, Y., S. Bengio, Y. Pouliot, and P. Agin (1990): A neural network to detect homologies in proteins. In: Touretzky, D. (ed.) Advances in Neural Information Processing Systems 2. Morgan-Kaufmann, pp. 423-430.
- Cai, D. and H. Zhao (2016) Neural word segmentation learning for Chinese. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp. 409-420.
- Castelvecchi, D. (2016): Can we open the blackbox of AI. Nature 538: 20-23.
- Hladka, B. and M. Holub (2015 A gentle introduction to machine learning for natural language processing: how to start in 16 practical steps. Lang. Linguist. Compass 9.2: 55-76.
- King, D. (2016) Evaluating sequence alignment for learning inflectional morphology. In: Proceedings of the 14th Annual SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pp. 49–53.
- Leslie, C., E. Eskin, A. Cohen, J. Weston, and W. Noble (2004) Mismatch string kernels for discriminative protein classification. Bioinformatics 20.4: 467-476.
- Mahendran, A. and A. Vedaldi (2015) Understanding deep image representations by inverting them. In: 2015 IEEE conference on computer vision and pattern recognition (CVPR), pp. 5188-5196.