Drawbacks of Deep Learning

In order to solve a problem, deep learning enables machines to mirror the human brain by making use of artificial neural networks. These networks are known to run a variety of applications such as speech recognition devices like Siri and Neuro-Linguistic Programming. Deep learning has hence been recognized as one of the major research areas required to advance AI.


Although the importance of deep learning is increasing and several advances in its research are touching great heights, there are a few downsides or challenges that have to be tackled to develop it. To begin, copious amounts of data are required to train deep learning algorithms – as they learn progressively. To exemplify, for a speech recognition program, data formulating multiple dialects, demographics and time scales is required to obtain desired results [1]. While firms like Google and Microsoft are able to gather and have abundant data, small firms with good ideas may not be able to do so. Additionally, data availability for some sectors may be sparse and thus hamper deep learning in that industry. For instance, health care where AI is used to recognize tumors in X-ray scans [2].


Moving on, though deep learning models are very efficient and are able to formulate an adequate solution to a particular problem once trained with data, they are unable to do so for a similar problem and require retraining [3]. To elaborate, these neural network architectures are highly specialized to a specific domain and reassessment is needed to solve issues that do not pertain to that identical domain. For example, Google’s DeepMind trained a system to beat 49 Atari games; however, each time the system beat a game, it had to be retrained to beat the next one [2]. Interestingly, while these algorithms did a great job of mapping inputs to outputs they were unable to understand the context of the data they were trained with. Designers of the algorithm claimed that the best way to win the game was to dig a tunnel in the wall after 240 minutes; nevertheless learning through multiple trials and errors the system was able to decipher this, but it was not aware of what a tunnel or a wall was [3].


Another challenge of deep learning is it requires large amounts of processing power. This high-performance hardware is mostly the multi-core high performing graphics processing unit or a similar processing system [1]. These processing units require and consume a lot of power and are therefore a costly affair.


To conclude, in the words of Andrew Ng, deep learning is a great way to “build an AI-powered society” [4] and I believe overcoming these shortcomings with the help of other technologies is the right way to achieve this goal.










[1] – https://hackernoon.com/challenges-in-deep-learning-57bbf6e73bb


[2] – https://www.theverge.com/2016/10/10/13224930/ai-deep-learning-limitations-drawbacks


[3] – https://bdtechtalks.com/2018/02/27/limits-challenges-deep-learning-gary-marcus/


[4] – https://becominghuman.ai/what-i-learned-from-andrew-ngs-deep-learning-specialization-ccf94fea2a0f


One comment on “Drawbacks of Deep Learning”

  1. Tanu,
    Good article. I agree with you about the drawbacks of Deep Learning (DL) you pointed to.
    Just wanted to add following comments on 3 limitation points you reveled in your post:

    – Data: In this reference [1], the author said it well: “The biggest limitation of artificial intelligence is it’s only as smart as the data sets served”

    – Learning algorithm complexity: Our speaker in class 3, Craig Martel from Linkedin mentioned that the most used algorithm in AI is linear regression. Although, that seem as a simple algorithm, running DL based on such algorithm have limitations because the variables injected in the algorithm become large multi-dimensional regressions to solve.

    – Computing Hardware: As you mentioned, the GCPUs and alike processors have limitations. Our speaker from IBM in class 3 had touched the subject of quantum computation. That solution looks very promising for reducing computation time and complexity. That is a better future to reduce computation complexity needed by DL.



Comments are closed.