Search
Close this search box.
Search
Close this search box.

Deep Learning Examples Of Shortcomings

Deep Learning God Yann LeCun

Deep Learning Examples Of Shortcomings Deep learning (DL) is analogous to  induction, albeit based on historical data fitting. If we use DL to forecast stock trends, it has a high probability of losing money, since our regulated market is constantly evolving. History may repeat itself, but there must be some changes. For example, US stocks are currently facing a slump. Primarily because of the Federal Reserve’s plans to increase interest rates and the tensions in Ukraine and Russia. As a result, which cause flows in the money market. Under this situation, DL may not be an efficient tool because objects of verification are always a problem.

Firstly, DL relies on historic information.

It is hard to infer the meaning of a word you have not seen, whether it is a person or a model. If you have not heard of it, you probably do not know it. 

Secondly, these nouns sometimes have a time nature, such as COVID. You certainly never heard of COVID10 years ago, and consequently it is difficult to have them in the data and learn the model. Now that we have these kinds of nouns, will we retrain the model? It will be unreasonable because of the efficiency. We cannot finish learning, and new words will continue to come out. 

Thirdly, nouns have strict textual meaning. “LA LA Land” is a movie, but “XA XA Land” is not. DL can easily generalize and summarize vocabularies, but sometimes it is not appropriate for specific nouns. Unfortunately, after the deep learning system is trained, we cannot be sure how it makes decisions. Deep learning systems may be good at recognizing patterns of pixel distribution, but they cannot understand the meaning of patterns and the reasons behind them. 

Additionally, this system is easy to be misled. Just one small change can generate significant deviations in the predictions.  Therefore, DL may sometimes be irrational in stock trading.I think only using DL in stock trading to make decisions is an adventure: it can be advantageous, but sometimes it will lead to a biased result, which is especially concerning since trading should be cautious and subtle.

Aside from those discussed, two other major shortcomings of Deep Learning (DL) include:

1. A large number of examples are required to solve problems. 2. Computationally intense to train and deploy. For example, DL has a serious problem of low efficiency. “Let a kid learn to know a dog,” “we do not need to say ‘dog’ 10000 times, but it takes so many times for the deep learning system to learn ‘dog’. Therefore, for simple problems, there is no need to use deep learning.

From my view, the macro analysis of the market, including the development of large institutions that have an impact on the world process, local wars that control the distribution of resources, policies designated by various countries, and the rise of emerging technologies,  will affect the market to a certain extent. These are the fundamental reasons for determining the market trend. Then, after we grasp the overall situation, we can screen the specific stock and the targets in line with the general trend by using DL.


More specifically, it is hard to predict a “black swan event” such as COVID-19, but it is not hard to predict the monetary policy (quantitative easing) for the US. Under this situation, it is better to choose the stocks which are likely to have a higher performance in COVID, such as Amazon and Zoom by using DL to find a signal of stock trading.

Deep Learning Examples Of Shortcomings