Though AI has made significant progress in the recent past, it has its own limitations like all other past approaches to making machines intelligent. Let us look at some of those limitations:
Needs large labeled data for training and testing
Most of the success AI has achieved so far has been in supervised learning, where labeled data (historic data that maps input and correct output) is required. This is a very labor intensive and time-consuming exercise since these algorithms need massive amount of data. As an example, to train an autonomous driving car, we need to label millions of images, with boundaries for various objects within each image. Then the algorithm learns to identify various objects surrounding the car, including other cars and vehicles, pedestrians, traffic lights, other traffic signs such as STOP, Yield etc.
Reinforcement learning techniques are being used to reduce the dependency on this labeled data. DeepMind in their AlphaGo algorithm used a lot of labeled data, but AlphaZero, their subsequent program did not use any labeled data. It was just given the rules of the game and allowed to play the game by itself and learn. Hopefully such methods would succeed in solving problems that are being solved currently by supervised and unsupervised learning problems also in future.
In some cases, Generative Adversarial Networks (GANs) are being used to generate test data to compliment limited labeled data, since GANs can generate data of any distribution.
In the case of images, we can use simple methods to generate a lot more training and test images by rotating (rotational variance) or moving up, down, left or right (translational variance).
Output logic generated is Opaque or Block Box
In the case of neural networks (deep learning), interpreting the output rules is not very transparent. It does not enable a what-if or scenario analysis between input and output.
Of late, few techniques are being developed to get an understanding of what’s really happening in each of the layers of the network. TensorFlow has come up with Tensor Board that gives a lot of visibility into how various training parameters across the layers are behaving as the training progresses.
Potential bias in the output
Gender, ethnic, regional biases can creep into the output generated by AI algorithms, depending on how the training data is sampled, or whether it represents the whole population adequately. There are some neutralization techniques available to reduce the effect these biases. However, they are currently not mature enough to avoid biases completely.
Generalization of the model
AI models often work better on training and validation data, but often fail when they go live in production. This could be because training, validation and test sets came from different distributions, or because the live production data distribution was slightly different from the data distribution that the algorithm was trained on.
To reduce this impact, we need to ensure that training, validation and test data sets represent actual production data distribution adequately. Using regularization techniques can also help avoid over fitting or under fitting the training data.