While AI is again on trend with a lot of fanfare, the things that drive this trend are Large Language Models (LLMs) and GPT as an algorithm. In contrast, the AI ​​product that our research institute offers is financial index volatility analysis, which belongs to the field of time-series analysis, and as I wrote in my previous column, we implement convolutional neural networks (CNN) and LSTM for this purpose. And as long as looking at the current condition, the latter seems rather unpretentious than the former, I feel. There was a trend at one point where the ‘AI Roboviser’ service based on the latter algorithm was developed for individual investors in Japanese investment brokerage firms. However, to be honest, there were no people who gained their profits from it emerged. Though ‘AI Roboviser’ keeps being provided, it is, in fact, obscured when compared to the flow of generative AI, such as ChatGPT which has been sweeping the world.

However, I cannot help but feel that this is precisely where the ‘pitfall’ lies in the current discussion surrounding AI. It is because GPT and LSTM share the same origin, and what each algorithm ‘does’ is fundamentally the same (pattern matching. As this is a really important view, it will be discussed in another column). In this, ‘language generative AI’ that is developed based on the former has been propagated like it would dominate the whole human society. Even though some conscientious pioneers in AI research say, ‘That’s not true!’, it is frightening that this trend continues unabated within the mass media. As soon as I say, ‘It is impossible for generative AI to overwhelm humanity,’ I feel like I’ll be looked down upon, and it scares me.

Put simply, the problem here is this. When comparing GPT and LSTM, the general public speaks as if the former surpasses the latter, and the reality today is that people are likely to ask, ‘What? Are they still using LSTM?’ However, I have to say that it is too rough to simply compare the algorithms themselves and draw conclusions about their usefulness.

(Source: arutema47’s blog)

When tracing the origin of GPT, it reached the algorithm called Transformer. And this Transformer originated from the discovery and invention of Attention (attention mechanism), however, in fact, it is just a revised version of RNN (Recurrent Neural Network). Anyone who has studied AI seriously knows this crucial fact. The series of algorithms that could be called the RNN family share a common characteristic in that they remember the results of differentiating the loss function concerning the most recent input values ​​and use this as a reference when performing similar calculations that are the current task at hand. Attention (attention mechanism) narrowed down the point on where in structured algorithm needs to be paid attention. As a result, when a word is input into a sequence of natural languages, it enables one to infer the next word more precisely (That is why Google Translate has drastically improved since 2017).

However. It means, in short, that ‘overfitting’ for input value in reality is regarded as natural in the algorithm after. In natural language, it is acceptable in the case of inferring the flow of words in sentences. This is because there is no “trend and noise problem” as occurs in time series analysis. While it may seem like there are infinite options for a word B following a word A, this is often not the case when considering fitting with the overall context. For example, it is highly unlikely that someone would suddenly start talking about the end of humanity while talking about dinner. Therefore, we should just play the ‘◎’ ‘×’ game within the linguistic space, which is narrowed down by itself, and that is the substance of language-generative AI. Moreover, it is officially acknowledged that the Large Language Model (LLM) is revised by the human hand in the end under the name of ‘avoiding inappropriate political/discriminative expression’. Can it still be said ‘almighty AI’?

So, what will happen when trying to create a tool like Prometheus of our Institute with Transformer? Put simply, ‘the huge expectation will be betrayed splendidly’. In fact, I tried the Transformer-based coding in parallel with the Prometheus POC, however, it was completely useless in a point-of-future scenario in financial volatility. In a nutshell, there are　fewer ‘plays’, which means ‘generalisation’ has a problem. The result was a series of values ​​displayed on a very jagged chart, and the impression I got was that it was useless as a tool for augmenting human capabilities. After all, time series data itself structurally contains human-created ‘noise’ in addition to trends inherent to the time series itself. Even if the latter is removed through pre-processing, the problem of ‘noise’ will inevitably arise at the final stage of comparison with reality. This means that if the algorithm is made to be ‘overfitted’ like a Transformer, it will be completely useless in the real world.

The impression in the situation above is that the effect that the Structural Topic Model, which has been on a trend for these last few years, gave to the Topic Model, which has not been spoken in NLP. The Structural Topic Model is known to be used by many social scientists because it skilfully deals with the axis of time, enabling the chronological ‘interpretation’ that is unique to natural language and, in particular, social penomena. In short, when it comes to social phenomena, if there is no room for interpretation on the part of the observer, it is possible that the only conclusions reached will not quite fit the situation. Furthermore, the Structural Topic Model, which allows for the role of ‘interpreter,’ is ‘human-friendly,’ and has been a blessing to many social scientists who had thought that the advent of AI would ‘eat away their livelihoods.’

The reason for using the simple structure of CNN+LSTM (moreover, MCMC) in Prometheus is because of such situations. It is ‘human’s hands, eyes and brains’ in the end. It is impossible for AI to exist on its own. There always needs ‘meaning-making’ by human hands. The more ‘language generation AI’ such as ChatGPT emphasises the ‘humanness’ of the answers it gives, the more we realise that the human world is not actually precise and that without a moderate amount of fuzziness, we will not ultimately achieve true accuracy. That is the reality surrounding AI.

10^th August, 2023

Marunouchi, Tokyo

Written by Takeo Harada

(CEO/Global AI Strategist, IISIA)