Pascal Compilers, Voltage Thresholds and Vending Machines

Body

Last week, fellow IBM blogger Barry Whyte Barry pointed out that my recent post on [Cognitive University for Watson Systems SmartSeller] was my 1,000th blog post. After 10 years of blogging, I have reached the 1,000 mark!

(As IBM is focused on its transformation from a "Systems, Software and Services" company to a "Cognitive Solutions and Cloud Platform" company, it seems appropriate to highlight my 1,000 blog post on the concept of cognitive solutions.)

A lot of people ask me to explain what exactly does IBM mean by "cognitive", which is a fair question. Let's start with the [Dictionary definition]:

cognitive

1. of or relating to cognition; concerned with the act or process of knowing, perceiving, etc.
2. of or relating to the mental processes of perception, memory, judgment, and reasoning, as contrasted with emotional and volitional processes.

What exactly does IBM mean by Cognitive? IBM has taken this definition, and focused on four key strategic areas:

Understanding

In the summer of 1981, I spent a summer debugging a "Pascal" compiler at the University of Texas at Austin. I wasn't told that was what I was doing. Rather, I was tasked with writing sample Pascal programs that would demonstrate the features and capabilities of the language.

Every day, I would come up with a concept of a program, punch up the cards, run it through the CDC hopper, and verify that it would work properly. If I didn't have it working by lunch, I would take it to the "help desk", they would look it over, and tell me how to fix it after I got back.

Most of the time, it was a mistake in my software. A few times, however, it was a flaw in the compiler itself. My programs were basically test cases, and the Pascal Compiler development team was fixing or enhancing the compiler code every time I had a problem.

Compilers basically work by parsing the program text, looking for fixed keywords that are entered in a specifically prescribed order to make sense. Other keywords may represent data types, variables, constants or pre-defined macros.

But compilers are not cognitive. Cognitive solutions can understand natural language, and have to handle all the ambiguity of words not being in the correct order, or different words having different meanings.

Reason

As an Electrical Engineer, I had to take many classes on classical analog signal processing. In fact, all computers have some amount of analog components, where threshold processing is used to differentiate a zero (0) from a one (1).

For example, if a "zero" value was represented by 1 volt, and a "one" value by 5 volts, then you can set a threshold at 3 volts. Any voltage less than 3 would be considered a "zero" value, and anything 3 volts or greater a "one" value.

But threshold processing is not cognitive. Cognitive solutions also use thresholds, but their thresholds are dynamically determined, through advanced analytics and statistical mathematical models, and may adjust up and down as needed, based on machine learning over time.

Learning

IBM Research is proud to have developed the world's most advanced caching algorithms for its storage systems. Cache memory is very fast, but also very expensive, so offered in limited quantities. Caching algorithms decide which blocks of data should remain in cache, and which should be kicked out.

Ideally, a block in read cache would be kicked out precisely after the last time it was read, with little or no expectation for being read again anytime soon. Likewise, a block in write cache would be destaged to persistent storage precisely after the last time it was updated, with little or no expectation for being updated again anytime soon.

Traditional approach is "Least Recently Used" or [LRU]. Cache entries that were read recently or updated recently, would be placed on the top of the list, and the least referenced would be at the bottom of the list. When space is needed in cache, the entries at the bottom of the list would be kicked out.

IBM's [Adaptive Cache Algorithm outperforms LRU]. For example, on a workstation disk drive workload, at 16MB cache, LRU delivers a hit ratio of 4.24 percent while ARC achieves a hit ratio of 23.82 percent, and, for a SPC1 benchmark, at 4GB cache, LRU delivers a hit ratio of 9.19 percent while ARC achieves a hit ratio of 20 percent.

But caching algorithms, including IBM's Adaptive Cache, are not cognitive. These algorithms respond pragmatically based on the current state of the cache. Cognitive solutions learn, and improve with usage. This is often referred to as "Machine Learning".

Interaction

The human-computer interface (HCI) has much room for improvement in a variety of areas.

Take for example a snack vending machine. In college, we had assignments to simulate the computing logic of these. We had to interact with the buyer, receive coins entered into the slot--nickels, dimes and quarters representing 5, 10 and 25 cents--determine a total monetary balance, and then dispense snacks of various prices and return an appropriate amount of change, if any. There is even a [greedy algorithm] designed to optimize how the change is returned.

But vending machines are not cognitive. Like the caching algorithms, vending machines interact based on fixed programmatic logic, treating all buyers in the same manner. Cognitive solutions can interact with different users in different ways, customized to their needs, and these interactions can improve over time, based on machine learning.

IBM is exploring the use of Cognitive Solutions in a variety of different industries, from Healthcare to Retail, Financial Services to Manufacturing, and more.

technorati tags: IBM, Barry Whyte, Cognitive+University, cognitive computing, Pascal Compiler, CDC, Electrical Engineer, LRU, caching algoritm, Adaptive Cache, human-computer interface, HCI