It isn’t AI yet!

Today we hear about large organizations taking strides towards implementing AI in their businesses, but the level of learning exhibited in these solutions are very limited. Though the intent of these solutions is commendable, the approaches used by them are limited and don’t bring out the essence of intelligence or learning, as claimed.

If you take the example of Amazon Go which talks about bringing together sensor fusion, machine-vision, and deep learning algorithms to achieve the check-out less experience. Though the solution of offering check-out less experience is fabulous, attaching a deep-learning tag to it makes it pretentious. Detecting a customer which product was picked by the customer and which product was kept back did not need deep learning but can be arrived with a simple state function

If you analyze deep enough, the intelligence required to achieve this objective is highly static. Identifying a customer in a store and keeping track of which product is picked or returned can be detected using a simple active combination of states between data collected from the mobile app and the associated sensor.

Just to explain in a simpler manner, the customer who walks into the store is detected through sensors (beacons/RFID).  When the product is in place(rack), you can have the default combination as shown as the first combination. When the product is moved out of the zone, the product reading would change as it loses its association with the defined zone. A more clear explanation on how it’s done can be found here.

If the sensors are placed strategically and are in sync with the mobile app of the customer, you could detect customer’s browser behavior inside the store using the simple match function for available combinations. I really can’t see why deep learning is required to understand customer actions. If Amazon was attempting to predict what the customer might buy and realigning product display on the rack, then you might want to learn from past behavior to inform the racks to auto-organize the products within.

The other big exponent of deep learning is Google, which demonstrates its deep learning through its AlphaGo program, uses pattern detection and weighting associated patterns for predicting available moves. Though this deep learning exercise from Google shows promise with its accuracy, the same cannot be said of its Search product. The search product behaves like a plain information system with random relevance.

You could try a few questions yourself. Donning the role of a kid, I searched for “Can tortoises fly?”. I was shown a list of links that had information that simply was not in context. A natural language processing system with intelligence would throw back answers saying “No. Tortoises cannot fly because….” and probably list helpful links.

Further up, you can see Google search performing just the opposite for queries like “How many union territories in India excluding Andaman and Lakshadweep?”. The search result seems to show only results related to Andaman and Lakshadweep. This clearly shows there is no effort done to understand the simplest of context markers. Queries like “Excluding North and West, which are the other two directions? simply drove links out of context.

In all these scenarios, the search engine just lists all related links. You could say it is more of listing information rather than an intelligent answer provider. As far as I understand, knowledge tools which use NLP for processing machine translation of the natural language, maintain a holistic hierarchical data structure with absolute relationships, which helps in understanding the context hidden in a given sentence. Delving into this link would give a clearer picture of what goes into the making of a conversational machine.

Though we can understand that the AlphaGo is a smaller and a lesser dynamic program than the Google Search, the real learning demonstration comes in an environment where the given data parameters are more in number and have hidden subnets. In these environments, patterns are many and it requires weight computation at multiple nodes to accurately deduce the right pattern.

If we are to anticipate major breakthroughs using AI, we need to grow up beyond these toy stories and make a true effort to match thought and reasoning patterns exhibited by natural intelligence to come good with our AI goals.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s