What AI Technologies Can You Apply To Your Own Research? -Advice from a Computer Scientist-

Kameda H

Published on: 2021-09-26

Abstract

Since about ten years ago, Artificial Intelligence (AI) technology has been affirmatively and successfully applied to a wide range of fields, for example, image recognition processing. Especially, Deep Learning technology gives many outcomes and impacts to our society. I also adopt the AI algorithm GAN (Generative Adversarial Network, 2014) to build a new iPSC-derived cancer stem cells detecting system [1-3].?But in the field of Artificial Intelligent research there are still many other interesting technologies with much amount of potential for you to succeed. For example, AIs of Natural Language Processing (NLP), i.e., 1) Information Retrieval for web pages, patent document, and answers to questions, 2) Document Clustering including mail filtering, and recommendation, 3) Extraction for keywords, genre-oriented expressions, and relationship, and 4) Transformation including translation, summarizing, text production, dialogue production, and so on. Nowadays, among them, two attractive NLP AI technologies have emerged; Bert (Bidirectional Encoder Representations from Transformers, 2018) and GTP (GTP-2, 2019 and GTP-3, 2020) [4-6]. Bert is proposed by Google based on a new deep learning technology “Transformer,” which processes natural language text by far more accurately than previous AIs. GPT-2 and GPT-3 are also the same kind of powerful AIs and freely available, which are proposed by OpenAI group. With these AIs, you may build many kinds of useful tools for your research at a relatively low cost, for example, to classify document, to extract key knowledge, to translate, and so on. But deep learning AIs have some weak points. One of them is that AI cannot analyze and explain causality. That is to say, AI cannot tell us what a cause of the results is nor on what ground AI made such a decision. AI system is just a black box, i.e., we pay attention only to input and output. As you’d know, even in statistics there have not yet been any theory of causality. But scientists like you are strongly interested in causality.

Keywords

Artificial intelligence; Generative adversarial network; Natural language processing

Advice from a Computer Scientist

Luckily, some feasible solutions to it have been recently proposed. One of them is the explainable AI (XAI). Certain AI researchers proposed two algorithms LIME (Locally Interpretable Model-agnostic Explanations, 2016) and SHAP (SHapley Additive exPlanations, 2017) as the very first step to seek explainability in AI [7,8]. They analyze the inner works of deep learning systems to suggest some grounds why the AI produced such outputs to the inputs. On the other hand, statisticians adopt another approach, i.e., Statistical Causal Inference and Causal Discovery. One example is SAM (Structural Agnostic Modeling, 2018) [9]. SAM is implemented based on the GAN to explore causal relationships. They are now rapidly developing and freely published via Website, e.g. Github sites.

If you really want to know the causality of phenomena, one solution is SEM (Structural Equation Modeling) in statistics, and another one is ILP (Inductive logic programming) which is proposed as a machine learning methodology (1995) [10]. ILP system, for example, Progol, discovers new concept and relation between some concepts by reasoning deductively, inductively, and abductively. This methodology is very interesting but still has a serious weak point that it takes huge amount of running time. We have to wait a little while.

 Finally, I introduce you, especially young researchers, some useful software, when you build your AI systems by yourselves. If you have already some knowledge and skills of programming, using Python3 is the best solution as the first step. And try to use Docker system, too, which may change your research life better [11]. For people who have little amount of programming knowledge and skills, there are luckily some non-programming tools. For example, NNC (Neural Network Console, Sony) [12]. Search for more tools and frameworks fit to your research through Web sites, and try it anyway, so that you’ll get plenty of outcomes with them soon.

References

  1. Goodfellow IJ, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative Adversarial Nets.
  2. Aida S, Kameda H, Nishisako S, Kasai T, Sato A, Sugiyama T. Conditional generative adversarial networks to model iPSC-Derived cancer stem cells. J Adv Comput Intellgen Intelligent Inform. 2020; 24: 134-141.
  3. Aida S, Okugawa J, Fujisaka S, Kasai T, Kameda H, Sugiyama T. Deep learning of cancer stem cell morphology using conditional generative adversarial networks. Special Issue Appl Artificial Intelligence Med Res. 2020; 10: 931.
  4. Devlin J, Ming-Wei Chang, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 2018.
  5. GTP-2, OpenAI, Wikipedia.
  6. GTP-3, OpenAI, Github.
  7. Ribeiro MT, Singh S, Guestrin C. Why Should I Trust You? Explaining the Predictions of Any Classifier. 2016; 1135-1144.
  8.  
  9. Kalainathan D, Goudet O, Guyon I, Lopez-Paz D, Sebag M. Structural agnostic modeling: Adversarial learning of causal graphs. 2018.
  10. Muggleton S. Inverse entailment and progol. New Gen Comp J. 1995; 13: 245-286.
  11. Docker.
  12. Neural Network Console.