The Dip

On Monday, Pierre Vandergheynst asked on his twitter stream:

"Will sparsity ever solve an actual problem? 3 days to make up my mind ..."

It seems to me that we are in some sort of a dip as Seth Godin describes it. People ask whether compressed sensing has peaked or any of such nonsense. Here is Seth's summary of what the dip looks like:

....Every new project (or job, or hobby, or company) starts out exciting and fun. Then it gets harder and less fun, until it hits a low point-really hard, and not much fun at all. And then you find yourself asking if the goal is even worth the hassle. Maybe you're in a Dip-a temporary setback that will get better if you keep pushing. But maybe it's really a Cul-de-Sac, which will never get better, no matter how hard you try. What really sets superstars apart from everyone else is the ability to escape dead ends quickly, while staying focused and motivated when it really counts.

Winners quit fast, quit often, and quit without guilt-until they commit to beating the right Dip for the right reasons. In fact, winners seek out the Dip. They realize that the bigger the barrier, the bigger the reward for getting past it. If you can become number one in your niche, you'll get more than your fair share of profits, glory, and long-term security. Losers, on the other hand, fall into two basic traps. Either they fail to stick out the Dip-they get to the moment of truth and then give up-or they never even find the right Dip to conquer.

Some of you may feel that you are in that region of doubts. But let me give you some example as to why I think it is nonsense. Take for instance the case of hyperspectral imagers. Hyperspectral imagers are likely to drop in prices from one to two orders of magnitude as a result of implementing a CS approach. David Brady thinks it is about one order of magnitude drop, I am of the opinion that it is two. The reason you have never heard of hyperspectral imaging or that you cannot get to tinker with hyperspectral data or read a good paper on hyperspectral classification and AI related subject is because for you to do that, you need to buy yourself a $100,000 camera. A good many companies offering hyperspectral cameras as a product do not list the price of these items on their website some provide only leasing information  Make that price tag  10K$ or 1K$ and now we are talking about appealing to a much larger section of the population from hackers and makers to consumers. We are talking about creating a niche although I can only guess it might be a bigger niche than 50MP DSLR cameras aficionados that essentially resembles perfectly a low end disruption scenario as described by Clayton Christensen (from wikipedia)

Christensen distinguishes between "low-end disruption" which targets customers who do not need the full performance valued by customers at the high end of the market and "new-market disruption" which targets customers who have needs that were previously unserved by existing incumbents....In low-end disruption, the disruptor is focused initially on serving the least profitable customer, who is happy with a good enough product. This type of customer is not willing to pay premium for enhancements in product functionality. Once the disruptor has gained foot hold in this customer segment, it seeks to improve its profit margin. To get higher profit margins, the disruptor needs to enter the segment where the customer is willing to pay a little more for higher quality. To ensure this quality in its product, the disruptor needs to innovate. The incumbent will not do much to retain its share in a not so profitable segment, and will move up-market and focus on its more attractive customers. After a number of such encounters, the incumbent is squeezed into smaller markets than it was previously serving. And then finally the disruptive technology meets the demands of the most profitable segment and drives the established company out of the market.

"New market disruption" occurs when a product fits a new or emerging market segment that is not being served by existing incumbents in the industry..^[6]

The reason this technology is currently in the dip stems in part because it needs to be hardened. But where are the other niches you ask ? Like any new concepts, some will yield some implementable technologies some will die. But one thing is for sure, you cannot decide which one will die until you go through the process of hardening it. This process is treacherous for anybody involved because as Jerry Weinberg reminds us, the stakes are high. After you have invested all your time in a specific idea

The thought that disaster is impossible often leads to an unthinkable disaster.

You cannot let something sink if you have spent so much capital in it. In that process however, you could  discover that it answers questions that are not being asked or to entirely different problems. For instance, Compressed Genotyping, is clearly answering a question that is being asked at the individual level but is seldom asked at a group level. If the whole population were to be affected by Tay-Sachs, I am sure it would be a question that whole society would ask and there would be clear incentives for solving it ... now. But as nobody expects the Spanish inquisition, it is likely to be a bad bet to devise an hypothetical test procedures for a future unknown disease that could decimate our civilization. And this leads me to Nick Trefethen's maxim :

If the answer is highly sensitive to perturbations, you have probably asked the wrong question.

And so when I hear about other dips ? I am thinking and wonder if  beating the diffraction limit with nothing or solving crazy problems could fit the bill of a problem worth fighting for. But then, I get reminded that it doesn't matter really because sparsity is a proxy for power laws, and if your sensor cannot take advantage of that, then maybe you have the wrong sensor or the wrong approach. Like Forrest, you should run...