Leading thought in entrepreneurship from the London Business School community

AI and Machine Learning For New Startups - Part 1

Thinking of an AI startup? Or maybe you are leveraging AI for your existing startup. If you work in the tech industry, it is hard to spend a day without hearing about “AI”, “Machine Learning”, “Deep Learning”, or the like. As Eamonn Carey, (Managing Director of TechStars UK) says, “it [AI] feels like a hygiene factor nowadays” for startups.

But just how does deep learning work? And, what is AI capable of these days? Knowing the answers to these kinds of questions will help you be a more successful founder while better managing the expectations of customers, employees and investors in your new start-up.

In Part 1 of this series, we take a look at some of the underlying questions around how AI works on a basic level and how we can use it in a new generation of entrepreneurship.

Is AI like the human brain?

Old computing science literature used to draw parallels between the way machines learn to the way the human brain works. As AI continues to evolve, inspiration for its development continues to be derived from biology, making brain anatomy a decent explanatory tool for deep learning. Below is a picture of a neuron:

The human brain is made up of approximately 100 billion of these brain cells, all connected to one another. The larger branch-like structure of one brain cell (dendrites - on the left) is connected to multiple smaller branch-like structures (synapse - on the right) of other brain cells. So, whenever any brain activity occurs, the big branches on the left receive electric signals from numerous other smaller branches to which they connect. In turn, this brain cell fires off an electric signal of its own through its smaller branch onto the big branch of other brain cells.

This “neural network” is how we process thoughts in our minds. For example, when your eyes see a picture of a friend, the image gets translated into electrical signals throughout your own deep neural network and eventually outputs the name of your friend. But how does deep learning mimic this biology? Below is one node of a deep learning network.

Looks similar, doesn’t it? Clearly, some inspiration drawn from biology. The node takes a series of inputs (digitised data of some kind), makes a math function (activation function) to determine the nature of the data, and fires a corresponding numeric output (to more nodes). An individual node like the above alone does not hold much deterministic power, but a network of these nodes (such as that shown in the picture below) can do a decide job differentiating between cat and dog pictures. However, computationally the technology is still far from strength of the ~100 billion cells in the human brain.

How does AI recognize images?

Deep learning models, like all computer algorithms, work with digitised data. Let’s use the cat picture below as an example:

The original of this image is 300 by 600 pixels, stored on 3 color channels (Red Green Blue - like all digital images). The total number of pixels in the image is 180,000 (300 x 600), but a total of 540,000 values exist, as there are 3 layers of pixel values; 1 for each color channel (180,000 x 3).

The learning occurs when the 540,000 numerical inputs are fed into each input node of the deep learning network (most left column of nodes in the network diagram). Weights (similar to those used linear regression) are then calculated for each of the inputs at each node throughout the network, left to right. Finally, the output is compared to the expected result. If the predicted result is different from the expected result, the network then adjusts all the weights in model so the predicted result in the next iteration is closer to the expected result. Numerous iterations are simulated for numerous training datasets (think 10,000 photos of cats).

In the case of a cat image, the output layer (right most layer) calculates a probability between 0% and 100% - if the probability is >50%, then it is a cat! After many training pictures, accuracy tests and benchmarking exercises against human performance, this deep learning network should then be powerful enough to clarify photos of cats it has not seen before - handy right?

This simple example aims to give you a high-level idea of deep learning models and their application robustness. While we have strong advancements in computational power, we have yet to achieve the level of power offered by the ~100 billion cells of the human brain. Furthermore, exercises similar to simple cat picture classifications require a fairly large amount of data to be processed by networks in a systematic and rigorous way. For this reason, it seems as though general AI still sits a long distance into the future.

I hoped this article helped you gain a better understand some basics of deep learning models and assisted in growing your foundational thinking of what is, and isn't, possible for your AI startup. While the ideas illustrated here are fundamental, they can be abstracted to larger, more complex models to form the basis for some very successful AI startups like OnFido (AI driven identify verification, which swaps cat photos for photos of you).

In Part 2 of this series (coming soon), we will examine what the deeper implications are for your startup - the difference between applied AI versus general AI, industry verticals, and the ways in which the old “big data” plays a part in developing deep learning models for your startup.

Add new comment

This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.
Enter the characters shown in the image.
By submitting this form, you accept the Mollom privacy policy.